Only bother calculating AMSE when it's required

[sd2k]

The AMSE calculation is a fairly significant chunk of
the hot etscalc loop but it's just discarded for most
calls; we only need it when actually using the AMSE values.
By only calculating it when actually required we can speed up the
overall fit process by around 20-25%.

[sd2k]

/ci benchmark

[github-actions]

Benchmarks

---

Current:

fit
  Instructions:             8300423 (No change)
  L1 Accesses:             11277716 (No change)
  L2 Accesses:                  135 (No change)
  RAM Accesses:                 869 (No change)
  Estimated Cycles:        11308806 (No change)

auto_fit
  Instructions:            37198148 (No change)
  L1 Accesses:             50396663 (No change)
  L2 Accesses:                 2737 (No change)
  RAM Accesses:                1071 (No change)
  Estimated Cycles:        50447833 (No change)

bench_vic_elec_fit
  Instructions:           504547097 (+0.000003%)
  L1 Accesses:            561608919 (+0.000003%)
  L2 Accesses:               117869 (No change)
  RAM Accesses:               10080 (No change)
  Estimated Cycles:       562551064 (+0.000003%)

Previous:

fit
  Instructions:             8300423
  L1 Accesses:             11277716
  L2 Accesses:                  135
  RAM Accesses:                 869
  Estimated Cycles:        11308806

auto_fit
  Instructions:            37198148
  L1 Accesses:             50396663
  L2 Accesses:                 2737
  RAM Accesses:                1071
  Estimated Cycles:        50447833

bench_vic_elec_fit
  Instructions:           504547083
  L1 Accesses:            561608904
  L2 Accesses:               117869
  RAM Accesses:               10080
  Estimated Cycles:       562551049